SPE-177829-MS

Smart Proxy: an Innovative Reservoir Management Tool; Case Study of a Giant Mature Oilfield in the UAE Mohaghegh, S. D., Intelligent Solutions, Inc. & West Virginia University, Abdulla, F., and Abdou, M. ADCO, Gaskari, R., and Maysami, M., Intelligent Solutions, Inc.

Copyright 2015, Society of Petroleum Engineers This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition and Conference held in Abu Dhabi, UAE, 9–12 November 2015. This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

ABSTRACT

Reservoir management requires tools that can (a) provide fast track and accurate assessment of a large

variety of operations, while (b) are capable of quantifying uncertainties associated with management

decisions. Reservoir managers must be able to compare and contrast a large number of development

scenarios, while taking into account the uncertainties and risks involved with each scenario, in a relatively

short period of time. To achieving this important task with traditional technologies one must either

sacrifice the accuracy or the speed.

While numerical reservoir simulation models can provide the required accuracy, they fall short in

providing the required speed. On the other hand, reduced models (conventional proxy models that rely on

analytical solutions, simplified physics-based models or statistics-based response surfaces) can provide

fast output (speed) but fail to fulfill the required accuracy.

Surrogate Reservoir Model (SRM) is a “smart” proxy of the numerical reservoir simulation model. SRM

is developed to address this short coming in reservoir management. SRM takes advantage of the machine

learning and pattern recognition capabilities of Artificial Intelligence and Data Mining (AI&DM) in order

to “learn” and then accurately replicate the functionalities of the numerical reservoir simulation model.

Smart proxy (SRM) runs at very high speed such that a single run of the smart proxy takes only a fraction

of a second.

This paper presents highlights of development and application of a smart proxy (SRM) in the case of a

giant mature oilfields in the United Arab Emirates. The SRM is developed for a multi-million cell, highly

complex, naturally fractured, carbonate, numerical reservoir simulation model (developed using an

industry standard commercial numerical simulator) with more than 500 wellbores. The SRM is validated

through blind simulation runs and is used to plan filed development while honoring a number of

operational constraints (such as limits on FBHP, GOR and WC) required by reservoir management team.

SRM was used to increase field production without drilling new wells. This was accomplished by identify

the optimum choke size schedule for each well in order to maximize oil production while minimizing the

water cut.

2 SPE-177829-MS

ORGANIZATION OF THE ARTICLE

This paper has been organized in the following fashion. It starts with an introduction discussing the use

of fast and accurate proxy models as a reservoir management tool. Then traditional proxy models are

reviewed and Surrogate Reservoir Model (SRM) is introduced as a smart proxy for numerical reservoir

simulation models. Next some details regarding the field and the simulation model used in this project is

presented, followed by more details on the development (training and validation) of the SRM.

It is a fact that it would be next to impossible to comprehensively explain in detail the development of the

SRM in one technical article. In order to learn more about how smart proxies are developed is to refer to

a book that is scheduled for publication by SPE on “Data-Driven Reservoir Models”1. Although this book

is not directly about smart proxy models and SRM (the book covers in much details how to build

comprehensive empirical reservoir models using field measurements), it is quite relevant and provides

extensive details on how a Data-Driven Reservoir Model is developed from scratch (another book

specifically about Smart Proxy Models is currently under consideration).

Smart proxies are essentially data-driven reservoir models. The remaining of this paper is dedicated to

results achieved by the SRM from optimizing the production through dynamic choke setting to uncertainty

quantification using the SRM.

INTRODUCTION

Development of numerical reservoir simulation was a necessary response to an ever growing

understanding of the complexities associated with oil and gas production from hydrocarbon reservoirs.

While it had a modest beginning in terms of size and detail, it enjoyed a reasonably fast growth. The

computational requirements of numerical reservoir simulation made it a good candidate for mainframe

computers and later for Unix2-based systems. It was not until mid to late 1990s that a considerable increase

in computational power of personal computers opened new doors in the numerical reservoir simulation

industry. Major vendors of reservoir simulators started deploying their software applications on PCs and

since then, use of mainframes and Unix-based systems have been all but abandoned and almost all the

commercially available simulators are now PC-based.

As the computational power of personal desktop computers increased in accordance to the Moor’s Law3,

so did the level of details that reservoir engineers and geoscientists included in their numerical reservoir

simulation analysis. Given the fact that discrete calculus is the foundation of numerical reservoir

simulation and modeling4 the reservoir rock is divided into a large number of small cells and the collection

of all these small cells form the geological (geo-cellular) model. Fluid flow through the interconnected

small cells with the inclusion of wellbores as inner-boundary conditions and other rocks (with their fluid

content – if any) surrounding the hydrocarbon reservoir, as outer-boundary condition, form the essence of

what we know today as numerical reservoir simulation. Therefore, as the size of the reservoir increases,

and the size of the cells being used to model the reservoir decreases (to include more details on reservoir

and flow characteristics), the computational footprint of the reservoir simulation increases dramatically.

1 The book scheduled for publication in the first or second quarter of 2016. 2 UNIX is a family of multitasking, multiuser computer operating systems that derive from the original AT&T UNIX, developed in the 1970s at the Bell Labs research center by Ken Thompson, Dennis Ritchie, and others. 3 The observation made in 1965 by Gordon Moore, co-founder of Intel that the number of transistors per square inch on integrated circuits had doubled every year since the integrated circuit was invented. Moore predicted that this trend would continue for the foreseeable future. 4 Discrete calculus is used to turn the non-linear, second order, partial differential equation that governs the fluid flow in porous media into a set of linear, algebraic equations that are easily solved with well-known numerical techniques.

SPE-177829-MS 3

There has been a race between the computational power that is now offered through parallel processing

of multiple CPUs and the number of grid blocks (cells) that are used to model complex reservoirs. Most

of the numerical simulation models that are developed for real reservoirs include well above a million

cells. The numerical reservoir simulation model that is the subject of this article has an original geological

model that includes more than 21 million grid blocks that have been up-scaled to 2.5 million grid blocks

for the fluid flow model.

Even in a computational environment that include tens of parallel CPUs, a single run of such a simulation

model takes several hours. In some cases that unconventional hydrocarbon reservoirs are being modeled

(steam injections and naturally fractured reservoirs that include large number of induced fractures such as

shale) the computational time of numerical models are measured in days and weeks. In other words, to

achieve the accuracy and the precision that numerical reservoir simulation models provide, one must

sacrifices speed and timely response to queries.

With the descriptions provided above regarding the computational footprint of the numerical simulation

models, it is quite understandable why many reservoir managers do not have a favorable view of the utility

of the numerical reservoir simulation models as a reservoir management tool. A reservoir manager needs

to make decisions on a timely manner and for that she/he needs tools that possess two major

characteristics, accuracy and speed.

1. The reservoir management tool must be accurate. It goes without saying that the quality of the decisions

made by a reservoir manager is a function of the accuracy of the tool that is used to model the reservoir.

The accuracy, here is referred to the degree of detail that is involved in the analysis that consequently leaves

its mark on the generated response. For example, if Decline Curve Analysis is the tool that is being used in

order to make decisions, it is obvious that no reservoir characteristics and no operational constraints are

used during the decision making process. Therefore, the quality of the decisions are compromised as the

tools being used approximate the reality in order to achieve the required speed.

It is a widely accepted notion amongst overwhelming majority of the professionals in the industry that

numerical reservoir simulation, with all its short comings, is the most detailed and the most accurate tool

for modeling fluid flow in conventional reservoirs, today.

2. The reservoir management tool must be fast. Reservoir managers are required to make decisions for short-

term, medium-term, and long-term planning of a reservoir. Furthermore, reservoir managers need to make

decisions that take into account (a) multiple scenarios, as well as (b) uncertainties associated with many

parameters involved in hydrocarbon production. Therefore, reservoir management tools need to have a

small computation footprint such that the can accommodate thousands of model runs (scenarios) to be

assessed in a short time period (minutes or hours, not days and weeks). Large number of model runs is

required to search large solution spaces and to quantify uncertainties.

While fulfilling one of the above two requirements (accuracy), numerical simulation models fail to

accommodate the second requirement (speed) in order to be utilized as an effective reservoir management

tool. Traditional proxy models that are currently used in the industry (simplified physics-based – reduced

order - models or statistics-based response surfaces) provide the speed but fail to fulfill the accuracy

requirements. Smart proxies are developed for the purposes of making numerical reservoir simulation

models a viable reservoir management tool.

Smart proxies such as SRM are highly accurate (without sacrificing the physics that have been modeled

into the simulator or reducing the space and time resolution) and very fast. A single run of a smart proxy

such as a SRM only takes a few seconds while its accuracy in replicating the response of the numerical

simulation model is in higher 90 percentiles. Therefore, SRM as a smart proxy, fulfills both requirements

(speed and accuracy) of a reservoir management tool.

4 SPE-177829-MS

TRADITIONAL PROXY MODELING

In order to define and distinguish smart proxy model such as Surrogate Reservoir Model (SRM) from

traditional proxy models, it is necessary to first define and explain traditional proxy models and then

define and explain the smart proxy in contrast with the traditional ones. Traditional proxy models that are

currently used in the industry can be divided into two categories (a) statistics-based proxy models, and (b)

the reduced order or reduced physics proxy models.

The statistics-based proxy models deal with the responses that are generated from the numerical

simulation model. That is why they have been dubbed “Response Surfaces”. Response surface proxy

models require a large number of simulation runs (usually in hundreds, if not thousands) in order to capture

some of the essence of numerical model’s behavior, and to be useful, while smart proxies are trained using

a handful of reservoir simulation runs. The SRM that is being presented in this article was developed using

only six simulation runs5.

Traditional, statistics-based, proxy models suffer from the well-known problems that are associated with

statistics, especially when it is applied to problems with well-defined underlying physics. One of these

well-known problems is the issue of “correlation vs. causation”. In other words, simply because two

variable correlate, it does not mean that one is the cause of the other. An example that has been documented

to demonstrate this point refers to the rate of divorce in the state of Maine (in the United States of America)

between the years 2000 and 2009 that demonstrate almost a perfect correlation with per capita

consumption of margarine in the United States. Although there is a perfect correlation between the two

(Figure 1 shows a correlation of 0.993), it is highly doubtful that one has caused the other.

Figure 1. Good example of lack of relationship between correlation and causation: Divorce Rate in state of Maine highly correlates to per capita consumption of Margarine in the United States.

5 We fully realize that this small number of simulation run seems quite counter-intuitive for those involved in development of traditional proxy models. This is why in many occasions some well-known and well-respected reservoir engineers, reservoir modelers (and developers) have used the phrase “to-good-to-be-true” to describe SRM. Of course that is before they are exposed to all the details that goes into developing a SRM. Once they learn how a SRM is developed and see the results first hand, the notion usually changes from one of dismissal through disbelief, to one of admission of brilliance.

SPE-177829-MS 5

The other well-known problem with all statistics-based approaches is that they imposes a pre-defined

functional form (mostly polynomial) to the data that is being analyzed or modeled. Of course, one can test

a large number of pre-defined functional forms (such as linear, polynomial, exponential, etc.) to finally

find the best match. But what if the data representing the nature of a given complex problem does not lend

itself to a pre-determined functional form and it changes behavior multiple times? Response surfaces are

not known to be able to create well-defined and robust input-output relationship between the variables

that are crucial in a numerical simulation model and the model’s responses.

Recently, some new statistics-based proxy models have surfaced that use Principal Component Analysis

(PCA) as their core technology. Some recently published work have selected to use different flavors of

the Principal Component Analysis such as Proper Orthogonal Decomposition (PDO) and Polynomial

Chaos Expansion (PCE) in order to develop proxies of the numerical reservoir models (Cardoso 2010 -

Chen 2013 - Klie 2013 - He 2014). In the opinion of the authors6 these new techniques will ultimately

converge to the type of response surfaces that have been around for decades. In our opinion, this is due to

the fact that they are being developed within the same computational paradigm as the numerical simulation

model, and therefore, it is unlikely for them to provide major breakthrough in this arena.

Furthermore, most of these techniques have only been applied to academic problems with very small

number of wells. The real challenge will surface as they attempt to demonstrate the capabilities of these

techniques when they are used to build proxies for full field industry-based numerical models with

hundreds of wells and millions of grid blocks, similar to the numerical reservoir model that is the subject

of this article.

The second class of proxy models that are quite popular are Reduced Order Models (ROM). Engineers

and scientists have invented many clever ways of reducing the order of the numerical simulation models

in order to overcome the long computational overhead. But of course, as long as one is operating within

the realm of a given paradigm, no gain in computational time is without paying a price. The price that is

paid by the ROM is accuracy of the models.

There are two main ways to reduce the order of a model7. One way is to reduce the resolution (both in

time and in space, but mostly in space). In this approach the geological model is grossly up-scaled, so

much that in some cases the solution for each well approaches the analytical solution with all its short-

comings. Workflows have been developed to increase the static resolution (resolution in space) from a

very coarse set of grids in steps in order to find the best middle ground (Williams 2004).

Another way of developing ROM concentrates on the physics of the problem rather than the space and

time resolution of the numerical solution. In this second approach physics of the model is reduced in order

to circumvent the computational time. Many of the most recent examples of such ROM approaches have

been applied to numerical modeling of production from shale (Wilson 2012). For example instead of

naturally fracture medium (dual porosity) that increases the computational overhead, extensively, some

have opted to use single porosity models that is adjusted to act like a dual porosity system. Similar

approaches have been adopted for dual permeability models. In these approaches physical proxies are

used to substitute the current understanding of the detailed physics. The final result of ROM approaches

is that a different problem is being solved, and not the one that originally was the intent of the numerical

simulation model.

6 We investigated, and practiced extensively with, PCA before concluding that we need to move on to another set of techniques that resulted in developing SRM. 7 All other traditional reduced models can eventually be classified as one of these two main categories.

6 SPE-177829-MS

SMART PROXY MODELING

Smart proxy modeling is a paradigm shift in developing proxies for numerical simulation models. It brings

“Big Data” solutions to the realm of numerical simulation. Smart proxies are developed using pattern

recognition and machine learning. They learn and then accurately mimic the simulator’s behavior with

lightning speed. Domain expertise determine what needs to be learned and then extracts the required

information from the simulation runs while expertise in artificial intelligence and data mining guides the

architecture and the topology of the learning process.

Surrogate Reservoir Models8 (SRM) are the only type of “smart” proxy that has been introduced in the

oil and gas industry. SRM takes a completely different approach to building proxy models when compared

to the traditional techniques. In this approach the model is not reduced (neither the physics nor the space-

time resolution) like ROM and pre-defined functional forms are not used for its development, like response

surfaces. SRM learns and then mimics the detail behavior of a numerical reservoir simulation model with

high accuracy and does it at high speed.

SRM is trained using machine learning and pattern recognition techniques. It is an integration of multiple-

specialized neural networks as well as a series of knowledge-based intelligent agents that work in concert

in order to achieve a set of predetermined objectives. SRM is the best example of data-knowledge fusion

at the forefront of advance data-driven analytics. Since SRM is NOT a statistical representation of the

simulation model, without domain expertise in reservoir engineering and reservoir modeling, developing

a technology such as Surrogate Reservoir Model (SRM) would have not been plausible.

The data that is used to train the SRM is extracted from the numerical simulation model runs and therefore,

the physics, and the resolution in time and space of the original simulation model is preserved. Since SRM

conforms to the system theory, it has an Input-System-Output topology and therefore is not a statistical

best-fit of the simulator responses. Using the characteristics of the geological model as well as the

characteristics of the flow model, and including the boundary conditions and the operational constraints

that are used in the numerical simulator as input, and coupling them with the corresponding simulator

output, a comprehensive spatio-temporal dataset is generated that includes details of fluid flow in the

given reservoir that the SRM needs to learn from. The spatio-temporal database must include everything

that one wished to teach a SRM.

A BRIEF EXPLANATION OF HOW SRMS ARE BUILT

As shown in Figure 2 every grid block from every run of the numerical reservoir simulation model

generates a record that is used to train the SRM. In this figure the grid block shown in pink (and marked

as Tier 0) is the target grid block. In every run, eventually every grid block in the numerical model becomes

the target grid block. For example given the fact that six simulation runs were used to develop the SRM

for this project, and that the model includes 2.5 million grid blocks, every single neural network is trained

using a database that included 6 x 2.5 = 15 million records at every time step. If we are considering about

100 time steps (a conservative number) for each of the runs, then the number of records (information

extracted from the numerical model used for training the smart proxy) that are available for training the

smart proxy for this study will be 100 x 15 million = 1.5 billion records.

8 It is important to know that several papers have recently been published using the term “Surrogate Model” (even those referenced in this paper). However, these papers seem to be using the term “Surrogate Model” only as a substitute for the term “Proxy Model”. Their “Surrogate Model” are NOT the same as the smart proxies that are the subject of this paper.

SPE-177829-MS 7

The information extracted from the numerical simulation runs that is embedded in this 1.5 billion records,

summarizes the pressure and saturation changes in each grid block as a function of all the surrounding

grid blocks (as shown in Figure 2) and all other inner and outer boundary characteristics of the system as

related to this specific grid block. This information is used to assimilate the comprehensive spatio-

temporal database that includes all that is needed to train a series of specialized neural networks.

SRM is validated using blind simulation runs in order to confirm its robustness and accuracy. This process

may take a few iteration for a beginner, but a good reservoir engineer and reservoir modeler will soon

become a capable SRM developer once they understand the mechanism through which machine learning

works. The developed and validated SRM is used for reservoir management and planning purposes.

Results of the final SRM can always be compared with results of the numerical simulator for a final run

to make sure that the SRM is realistically and accurately reproducing the numerical simulation result.

The abovementioned steps were implemented during the development and validation of the SRM that is

presented in this study. SRM is defined as: an ensemble of multiple machine learning technologies

(intelligent agents as well as neural networks) that are trained to learn and then accurately mimic the

intricacies and nuances of the physics of fluid flow in a given hydrocarbon reservoir using (input and

output) data generated from the numerical simulation model.

The trained and completed SRM has a small computational footprint, such that thousands of SRM runs

can be made in seconds or minutes. This allows examination of a massive number of scenarios in a short

period of time, therefore making it practical to exhaustively search large solution spaces for optimal or

near optimal solutions, or to perform Monte Carlo simulations to quantify uncertainties associated with

geological models that forms the foundation of the numerical flow models.

Figure 2. Every grid block in the model is used as a training record. All the neighboring grid blocks contribute to the model response of the target grid block. Every time step in a single run of the numerical simulation model that contains 2.5 million grid block generates 2.5

million records worth of information for training.

8 SPE-177829-MS

THE MATURE FIELD IN THE UAE

The main objective of this project was to identify and rank the wells in this giant mature field that would

most benefit from a rate relaxation program. This filed is producing from two reservoirs that are referred

to (in this article) as Reservoir A and Reservoir B. There is a set of guidelines established by the

shareholders that emphasizes a total production cap on the field (300,000 BOPD) and a cap on oil

production (1,500 BOPD for the reservoir A and 2,500 BOPD for the reservoir B). Furthermore, the

guidelines that must be followed emphasizes that the Flowing Bottom-Hole Pressure must always remain

at least 200 psi above the bubble point pressure, and the Gas Oil Ratio at each individual well may not

increase above 3000 scf/stb.

This giant mature field includes more than 15 billion barrel of reserves. Production from this field started

in 1962 with water injection starting in 1974 into the peripheral aquifer. More water injection in the mid-

section of the reservoir started in 1994. A gas injection program in the field also started in 1992. There is

a facility back pressure of 500 psi on all the wells. The GOR and the water cut in the field at time of this

project was 1,350 scf/stb and 17%, respectively.

The geologic model developed using a commercial software application included more than 21 million

grid blocks. The geologic model was up-scaled to almost 2.5 million grid blocks to be used in the flow

model. The model included two different PVT regions and 8 different rock types, 12 regions for the fluid-

in-place, multiple high permeability streaks, with the typical inter-well spacing of about 1 kilometer.

Relative location of more than 500 wells that were modeled in this project are shown in Figure 3.

Figure 3. Well locations in the giant mature field in the UAE.

SPE-177829-MS 9

A single run of the numerical reservoir simulation model (developed and history matched by the operating

company using a commercial simulator) for this giant mature field in the United Arab Emirates takes

several hours using a large cluster of multiple CPUs. Achieving the objectives of this project required

searching a very large solution space. To realize the massive scale of the solution space that should be

searched for the optimum or near optimum solution, one must account for the fact that the wells

communicate with one another and their production scheme impacts production (and water cut) of the

neighboring wells (offsets). This is a dynamic problem that is being continuously modified. Furthermore,

the geological model that is forming the foundation of the flow model includes large amounts of

uncertainties that need to be quantified during the decision making process.

The facts numerated above makes performing such analyses impractical, even if the numerical model

would take only 20 minutes per run, let alone the hours that are needed for a single run of this numerical

model for this field. Assuming a highly unusual 20 minutes per run, requiring a conservative 10 million

simulation runs to achieve the objectives of this project (brute force) translate to more than 380 years of

only computational time. The impracticality of such scales are readily evident. That is why numerical

reservoir simulation models are hardly ever considered as serious reservoir management tools. Of course

there are experience modeler and engineers that can considerably reduce the required computational time

by carefully analyzing each simulation run in detail before prescribing the next set of runs and repeat this

process until they can reach a “good” solution. But how many of such engineers and scientists are

dedicated to each project? The impracticality of using numerical simulation models for such reservoir

management analysis is hardly ever disputed.

The SRM developed for this project that is presented in this manuscript achieved this objective with brute

force of making millions of runs and analyzing its results using data-driven analytics and generating a

dynamic FBHP schedule for each individual well to achieve highest possible oil rate while never violating

the guidelines put forward by the shareholders.

This project was commissioned after the results from a similar previous project on a neighboring field was

comprehensively analyzed by the operating company. In that project all predictions made by the SRM

was validated to be true in a study performed many years after the completion of the study, and examining

actual field operation results. Details of that project that resulted in $1.9 Billion dollars of incremental oil

was published previously (Mohaghegh 2014).

DEVELOPMENT AND VALIDATION OF THE SRM

As it was mentioned earlier, one of the major advantages of SRM is that its development requires only a

small number of simulation runs. This may sound counter-intuitive to those that are familiar with geo-

statistics and its use in developing proxy models. The reason that small number of simulation runs are

enough to develop a SRM can be explained through the use of the information generated by a single

simulation run and how it is used to train a SRM. Here we attempt to shed some light on this issues.

The first distinction that needs to be drawn between SRM and the traditional proxy models is the

philosophy and the approach that is used in developing proxy models. In traditional statistics-based proxy

modeling (response surfaces) the objective is to use some type of pre-determined functional form in order

to curve fit a series of inputs (usually reservoir characteristics and some operational constraints that have

been combined using several techniques in order to represent the state of the reservoir and the operation)

to a few output (usually simulator’s response such as oil production). In SRM the objective is not curve

fitting. The objective is to “teach” reservoir engineering to a machine (to a computer program in this case).

Since the statement “teaching reservoir engineering to a machine” may sound a bit too general, let’s break

it down to its components.

10 SPE-177829-MS

We can make the observation that a numerical reservoir simulation developed for a specific reservoir is a

collection of grid blocks as shown in Figure 2. Given the fact that there are 2.5 million grid blocks in the

model and that production from the wells (which is really the result of the change in pressure and saturation

at every grid block) is calculated at every time step (assuming 100 time-steps) and that we have made six

simulation runs for the development of this particular SRM, the number of records that is the same as the

number of examples to learn from becomes 2.5 million x 100 x 6 = 1.5 billion records. In other words we

have 1.5 billion examples of changes of pressure and saturation in a grid block to learn from. Furthermore,

heterogeneity of a reservoir would contribute to diversity and differences between static properties of each

grid block (and its associated neighboring grid blocks). This diversity helps the learning (training) process.

Although there may be a degree of redundancy (from a mathematical point of view) that is occurring, it is

a good thing when learning by example is taking place.

We think everyone would agree that “any” type of changes in pressure and saturation that can possibly

take place in any kind of run in this reservoir simulation model (while using it to perform optimization or

uncertainty quantification) or something close to it has already happened during this 1.5 billion examples.

This is actually the essence of smart proxy modeling and the SRM. If this can be accepted, and understood,

then the rest is details and what remains to be explained is actually how to perform the training. Implement

this not-so-trivial task (training and deployment of the SRM) and the nuances associated with it will

become a matter experience.

The objective of this project, as was mentioned earlier, was to see if more oil can be produced from the

existing wells (though saving the capital expenses of drilling new wells) by optimizing choke settings on

each well without violating the shareholder’s guidelines or drilling new wells. This project included

multiple SRMs specialized for specific purposes and specific scenarios. Here we only present one of the

SRMs that was set to accomplish the abovementioned objective while not increasing the total field

production of 300,000 barrels per day and the total planned water and gas injection. To develop the SRM

for this specific scenario we designed six simulation runs. The six simulation runs were used to train and

calibrate the SRM and a 7th simulation run was designed to serve as the blind run to validate the SRM.

Figure 4. Change in production cap for each individual well as a function of time. These two production schemes were designed as used for the training purposes of the SRM.

The six simulation runs used to train the SRM differed in the production cap that was imposed on each

individual well. The production caps imposed during four of the simulation runs that was kept constant

during the 30 years of predictive simulation runs were 1,500 – 2,500 – 3,500 – 4,500 barrels per day,

SPE-177829-MS 11

respectively. Two more simulation runs were designed, during which the production cap would change as

a function of time. Since we expected to change the choke settings (translating to change in the FBHP) in

order to achieve optimum production from each well, we needed to train the SRM such that it has seen

occasions where the production cap in the individual wells does not remain constant and is subject to

change over time.

Figure 5. Demonstrating the accuracy of the SRM. Figures show the SRM results (one of the six runs used for training and calibrating the SRM) against actual simulation runs - 1500 BOPD production cap for individual wells. Results from individual wells are summed in order to

calculate the results for the entire field.

The two production schemes of step-change (where production cap on each individual well increases by

500 barrels per day every few years) and continuous change (where production cap on each individual

well increases slightly every month) as shown in Figure 4. These production schemes were used as the

12 SPE-177829-MS

simulation runs number five and six. In the seventh simulation run that was designed as a blind run for the

SRM validation all wells were producing against an operational constraint that was translated to a

production cap of 3,000 barrels per day.

Upon completion of the SRM we needed to check its accuracy. First we compared the results from the

SRM with one of the runs that was used in the training. The comparison was made on a well by well basis

and also when all the results from all the wells were summed to represent the production from the entire

field. Figure 5 shows the results of this test for the entire field. This figure includes three graphs. The

graph on the top shows the results of the oil production for the entire field. In this graph the annual oil

production rate results generated by the SRM are shown by solid line while similar results from the

numerical simulation model are shown using green circles. Furthermore, the cumulative oil production

from SRM and the numerical simulation model is shown in different shades of green. Since the SRM

result are very accurate, no differences in the cumulative oil productions can be detected in this figure.

The two bottom graphs (in Figure 5) show total field water cut (left) and GOR (right) as a function of

time. In these figures, like the one before, the results generated by the SRM are shown by solid line while

similar results from the numerical simulation model are shown using circles (dots). Graphs displayed in

Figure 5 clearly show that the SRM has been trained well and provides results similar to the numerical

reservoir simulation model with high accuracy. The SRM for this field that its results are shown in Figure

5 runs at an unprecedented speed of only seconds for a complete run on a common desktop PC.

Figure 6. SRM's blind validation test for Well #17. Results from the SRM (solid lines) and numerical simulation model (dots) for FBHP, GOR,

WC and Oil rate versus time.

SPE-177829-MS 13

Figure 7. SRM's blind validation test for Well #384. Results from the SRM (solid lines) and numerical simulation model (dots) for FBHP,

GOR, WC and Oil rate versus time.

Figure 8. SRM's blind validation test for Well #223. Results from the SRM (solid lines) and numerical simulation model (dots) for FBHP,

GOR, WC and Oil rate versus time.

14 SPE-177829-MS

As was mentioned earlier, the results of the SRM are generated for each well individually. To show the

results of the SRM on a well by well basis Figure 6 through Figure 8 are presented. These figures represent

the accuracy of the SRM in generating results for the operational condition that it has not explicitly been

trained on, also known as the blind test. These figures show samples of the results of the seventh simulation

run that was put aside to test the robustness of the SRM.

Figure 9. Demonstrating the accuracy of the SRM. Figures show the SRM results against the blind simulation run. Results from individual wells are summed in order to calculate the results for the entire field.

Each of the figures (Figure 6 through Figure 8) include four graphs. In all the graphs the results generated

by the SRM is shown by solid lines while similar results from the numerical simulation model are shown

using circles (dots). For each well Oil rate, Flowing Bottom-Hole Pressure, Gas Oil Ratio, and Water Cut

SPE-177829-MS 15

are generated by the SRM and are compared with the results from the numerical simulation model. The

accuracy of the SRM in generating Oil rate, Flowing Bottom-Hole Pressure, Gas Oil Ratio, and Water Cut

in new operational conditions (as far as the training runs for the SRM were concerned) are clearly

demonstrated in these figures.

Figure 9 displays the results of the Oil rate, Gas Oil Ratio, and Water Cut for the entire field (for the blind

simulation run) when the results generated by the SRM (solid lines) are compared with the similar results

from the numerical simulation model (circles - dots). With the results of the validation run as displayed in

Figure 6 through Figure 9 development of the SRM is judged to be satisfactory and the SRM can now be

used for the purposes of optimizing the oil production from this field.

OPTIMUM PRODUCTION SCHEDULING

Once the SRM is developed and validated through the blind simulation run, it is time to use it to

accomplish the objectives of the project. The objective of the study was to (a) develop an optimized choke

setting schedule for each well that would maximize oil production (from its current cap of 1,500 BOPD

for the reservoir A and 2,500 BOPD for the reservoir B) without violating the shareholder’s guidelines

regarding FBHP and GOR, and (b) to rank the wells in the reservoir based on their potential of incremental

oil production.

Figure 10. Flow chart for optimizing the oil production rate from each individual well.

Fluid flow through a naturally fractured and geologically complex porous media is a dynamic and non-

linear phenomenon. The complex nature of this phenomenon manifests itself in the intricate behavior of

16 SPE-177829-MS

oil and water production from each individual well as the operational constraints are modified in the well

and its offsets.

Figure 11. Oil production optimization of Well #78 using the SRM. “Original” versus “Optimized”. The graph on the bottom demonstrates that all shareholder’s guidelines are honored during the optimization process.

Therefore, one should expect that a specific (constant) choke setting (production cap) would NOT be

sufficient for extracting optimum production from a well. In other words, a dynamic choke setting

SPE-177829-MS 17

(production cap) must be expected for optimum performance9 of individual wells. Furthermore, it is

conceivable that such a choke setting program (shown here as a production cap that is imposed on each

individual well) would be a function of pressure and fluid saturations around each individual well. Values

of these parameters are constantly changing as a function of production and injection throughout the

reservoir. This simply means that the optimum choke setting (imposed production cap) must naturally be

a function of time, as well.

Figure 12. Oil production optimization of Well #200 using the SRM. “Original” versus “Optimized”. The graph on the bottom demonstrates

that all shareholder’s guidelines are honored during the optimization process.

9 The optimum performance is defined as a series of choke settings (production caps) that would change as a function of time (annually) to provide maximum oil production while maintaining an acceptable level of water cut and honoring shareholder’s guidelines on FBHP and GOR.

18 SPE-177829-MS

In order to accomplish the objectives of this study, as was mentioned above, the model should be executed

in a matter that would accommodate these dynamics. Figure 10 is the schematic diagram of the workflow

that was used to implement the optimization for this study. This flow chart clearly demonstrate the type

of challenges that a study like this imposes on the model being used as the objective function of an

optimization routine.

The dynamic nature of the problem requires that only one time step of a complete simulation run be useful

at a time and the rest are discarded. The result of each run is used to determine the set of input values

(operational constraints) for the next run that would represent the next time step. The original run that is

performed and is shown using dashed lines in Figure 11 through Figure 14 demonstrates that each well

may or may not be able to produce even the target production cap. The objective is to identify the set of

choke settings that maximizes this production.

During the optimization routine using the SRM as the objective function to maximize the oil rate while

honoring the shareholder’s guidelines, each time step starts with the original production cap for the well

(1,500 BOPD for the reservoir A and 2,500 BOPD for the reservoir B). If the well is capable of supporting

the identified production cap without violating the guidelines, then the production cap is increased to the

next level and the process is repeated. This process is continued until the increase in production is no

longer sustainable (for this time step) and production no longer can be increased.

If the model decides that the production cap cannot be met without violating the guidelines, then the

process reverts to the original plan, or even lower production rates, such that the guidelines remain intact.

The fact that this process should be executed for each of the wells individually, demonstrates the number

of times that the objective function (which is the SRM in our case and would have been the numerical

simulation model in the absence of SRM) must be executed. Running this optimization algorithms with

the full field numerical simulation model would be impractical since it would take years to complete.

Therefore, during the designed optimization routine presented here an optimum choke setting schedule is

developed for each individual well.

Again, it is important to note that during this search and optimization process the guidelines put forward

by the shareholders must not be violated. In other words, the optimum choke setting schedule for an

individual well must honor the imposed guidelines. The characteristics of the optimum choke setting

schedule are as follows:

1. It results in maximum incremental oil production when compared to the constant, original

production cap schedule identified in the “original” in Figure 11 through Figure 14.

2. It honors a water cut maximum of 50%10.

3. It honors the shareholders’ guideline on FBHP (200 psi above the bubble point).

4. It honors the shareholders’ guideline on maximum allowable GOR (3000 scf/stb).

Figure 11 through Figure 13 show the results of the optimization process for three individual wells as

examples. Each of these figures include four graphs. Figure 11 shows the optimum choke setting schedule

(the red line in the bottom graph, identified as “Well Cap”) for Well #78 when compared to the “Original

Scenario” run.

The top graph in this figure shows the annual and cumulative oil production when original production cap

is imposed (blue dash line) along with the oil production (solid green line) that results from the optimum

10 Only when increasing the production cap. While producing at the original production cap the well is allowed to follow the water cut imposed by the simulation run (up to 90%).

SPE-177829-MS 19

choke setting (a.k.a. variable production cap corresponding to the red line, bottom graph). The change in

the cumulative oil production for this well is shown in the top graph (shaded areas).

Figure 13. Oil production optimization of Well #536 using the SRM. “Original” versus “Optimized”. The graph on the bottom demonstrates

that all shareholder’s guidelines are honored during the optimization process.

20 SPE-177829-MS

Figure 14. Oil production optimization for the entire field using the SRM. “Original” versus “Optimized”. The graph on the bottom

demonstrates that all shareholder’s guidelines are honored during the optimization process.

The middle graphs depict the comparison of GOR (second top graph) and WC (second bottom graph) for

optimized setting (solid green line) and “Original Scenario” (Dashed line). The lines and the shaded areas

in the bottom graph show the adherence of this new optimum well cap schedule to the shareholder’s

guidelines. The Flowing Bottom-Hole Pressure is shown with black line (it always stays higher than 200

psi above the bubble point). The GOR is shown in brown shaded area, maintaining a value below the 3000

SPE-177829-MS 21

scf/stb, and finally the blue shaded area shows the water cut. Figure 12 and Figure 13 show similar analysis

for Wells #200 and #536.

Performing the choke optimization process for all the wells in the field and summing the results provide

a picture of the overall field performance during the analysis period. Figure 14 shows the full field results.

This figure indicates that a large amount of gain (millions of barrels) can result from optimizing the choke

settings (developing customized production cap settings for individual wells based on their production

capabilities) in this field. Furthermore, this figure shows that the above gain (more than tens of millions

of barrels) in oil production is achieveable with minimal increase in Water Cut (less than 3% over the

same period) and Gas Oil Ratio (which will always remain below 2.0 Mscf/stb).

CANDIDATE SELECTION FOR RATE RELAXATION

During this step, wells are divided into four clusters based on their potential for producing higher

incremental oil without violating the shareholder’s guidelines. Cluster 1 wells are identified as “Primary

Candidate Wells”. These are the wells with highest expected incremental oil production upon modification

of choke settings (rate relaxation). Cluster 2 wells, identified as “Secondary Candidates,” will also benefit

from a rate relaxation program, albeit at a lower degree. Cluster 3 wells are identified as “May or May

Not be a Candidate”. Whether these wells benefit from a rate relaxation program may be contingent upon

more factors than those considered in this study therefore caution is recommended when selecting these

wells as candidates for rate relaxation. Finally, Cluster 4 wells are those identified as “Not a Candidate”.

These wells are not recommended to be rate relaxed. Figure 15 shows the locations of all candidates in

the field. The expected incremental oil production from the primary candidate wells are shown in Figure

16.

Figure 15. Map showing the location of all candidate wells for the rate relaxation program. Scenario LTP-FC

22 SPE-177829-MS

Figure 16. Incremental Oil production (till year 2020) for the “Primary Candidates” for rate relaxation program wells. LTP-FC.

UNCERTAINTY QUANTIFICATION

Geological models that form the basis of all numerical simulation models are uncertain. If the analyses

that are made based on the numerical reservoir simulation models are to be taken seriously, they must

include uncertainty quantification. Monte Carlo simulation is used to quantify the uncertainties associated

with the geological model.

The main reasons comprehensive uncertainty analyses are hardly ever performed with numerical reservoir

simulation models are:

a. Monte Carlo simulation requires that the objective function (numerical simulation model) be

executed hundreds (thousands) of time (to give the analysis statistical significance) for each step

of the analysis, and

b. In most cases, it takes tens of hours to make a single run of the numerical simulation models.

Therefore, making thousands of runs to quantify uncertainties associated with the geological model

for each possible scenario, is impractical.

Since Surrogate Reservoir Models (SRM) accurately reproduces the results of numerical simulation

models, they can be effectively used for uncertainty quantification. In this study, a comprehensive

uncertainty analysis is performed on all the wells that are identified as primary and secondary candidates

as well as those that may or may not be a candidate. Figure 17 and Figure 18 are examples of uncertainty

analysis performed on oil rate and water cut for two of the primary candidate wells. In each of these figures

the areas (yellow for oil and light blue for WC) shows the bounds of uncertainties associated with the

predictions made by the SRM, using the Monte Carlo Simulation method.

SPE-177829-MS 23

Figure 17. Quantifying the uncertainties associated with the geological model using Monte Carlo Simulation method; Well “320”.

Figure 18. Quantifying the uncertainties associated with the geological model using Monte Carlo Simulation method; Well “438”.

24 SPE-177829-MS

CONCLUSIONS

In this paper a smart proxy of a complex, full filed numerical reservoir simulation called “Surrogate

Reservoir Model-SRM” was presented. It is called “smart proxy” since it is fundamentally different from

the traditional statistics-, mathematics-, and physics-based proxy models. Smart proxies learn the

numerical model’s behavior through observation (training on the data generated by the numerical model)

and then mimic it accurately and at high speeds.

The SRM was developed using six simulation runs and was validated with one blind simulation run. It

was then used to optimize production from a giant mature field in the United Arab Emirates. The process

included hundreds of thousands of simulation (SRM) runs to be completed.

REFERENCES

Cardoso 2010: Linearized reduced-order models for subsurface flow simulation, Cardoso, M.A.,

Durlofsky, L.J., Department of Energy Resources Engineering, Stanford University, Stanford, CA 94305,

USA. Journal of Computational Physics 229 (2010) 681–700

Chen 2013: A Black-Box Stencil Interpolation Method to Accelerate Reservoir Simulations, Chen, H.,

Klie, H., Wang, Q., SPE 163614. SPE Reservoir Simulation Symposium held in The Woodlands, Texas,

USA, 18-20 February 2013.

He 2014: Reduced-Order Modeling for Compositional Simulation by Use of Trajectory Piecewise

Linearization, He, J., Durlofsky, L., SPE Journal 2014.

Klie 2013; Unlocking Fast Reservoir Predictions via Non-Intrusive Reduced Order Models, Klie, H., SPE

163584, SPE Reservoir Simulation Symposium held in The Woodlands, Texas, USA, 18-20 February

2013.

Mohaghegh 2014: Production Management Decision Analysis Using AI-Based Proxy Modeling of

Reservoir Simulations – A Look-Back Case Study. Mohaghegh, S. D., and Abdulla, F., SPE 170664. SPE

Annual Technical Conference and Exhibition. Amsterdam, the Netherlands, 27–29 October 2014.

Williams 2004: Top-Down Reservoir Modelling. Williams, G. J. J., Mansfield, M., MacDonald, D. G., &

Bush, M. D. (2004, January 1). Society of Petroleum Engineers. doi:10.2118/89974-MS.

Wilson 2012: Computational Optimization of Shale Resources Development using Reduced Physics

Surrogate Models. Wilson, K.C., Durlofsky, L.J., SPE 152946, SPE Western Regional Meeting,

Bakersfield, CA. 19-23 March 2012.